A Physics-Informed Neural Network Approach to Augmented Dynamics Visual Servoing of Multirotors

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Visual servoing enables robots to use camera feedback for precise motion control.
• Multirotor dynamics present challenges due to complex control inputs and potential uncertainties.
• Physics-informed neural networks (PINNs) offer a powerful tool for modeling complex systems with limited data.

Purpose of the Study:

• To develop a robust visual servoing strategy for multirotors by integrating PINNs with dynamics-centered control.
• To eliminate the need for inverse Jacobian calculations in multirotor motion control.
• To enhance the robustness of visual servoing against uncertainties in camera and multirotor parameters.

Main Methods:

• A physics-informed neural network (PINN) is employed to estimate system uncertainties and inaccuracies.
• The PINN model is integrated with a dynamics-centered visual servoing technique, directly mapping pixel variations to torque and thrust inputs.
• A nonlinear model predictive controller (NMPC) with an adaptive horizon is utilized for real-time implementation.

Main Results:

• The proposed method reduces the need for labeled data by 65% compared to existing data-driven approaches.
• The integrated system demonstrates robustness against up to 70% uncertainty in camera parameters.
• The NMPC enables control effort processing 10 times faster than conventional MPC strategies.

Conclusions:

• The combined PINN and dynamics-centered visual servoing strategy offers a robust and data-efficient solution for multirotor control.
• This approach effectively handles system uncertainties and modeling inaccuracies, crucial for real-world applications.
• The real-time capabilities of the NMPC ensure practical implementation for dynamic trajectory tracking.